The AI race is shifting from bigger models to cheaper, smarter systems

For the past two years, scoring points in the AI ​​race has been easy: bigger models, better benchmarks, and whatever company can claim the lead, at least until the next launch.

This scorecard is starting to look incomplete.

As companies move from testing AI to using it in real products and workflows, it is no longer a matter of arriving at the best model, but the one that is best suited for a particular job, at the right cost, with the necessary data, and in a chosen environment.

This shift opens the door to a new kind of AI competition that focuses less on model size and more on routing, cost, control and computing.

“The model alone is no longer the product,” Perplexity CEO Aravind Srinivas told CNBC. “It’s the harness, the orchestration system that puts the model into a very capable harness and pairs the model with many tools.”

This means that AI products become systems that can decide which model to use, when to use it, and what external tools or company data sources are required. A customer service role may not need the most expensive model. It could be a complex coding issue. A routine internal workflow can run on a cheaper open model. A more difficult step can be increased to a stronger step.

“The answer is always to use what is best for the task,” Srinivas said.

The emergence of alternative models comes as corporate America tightens its belt on AI spending and poses another challenge for OpenAI and Anthropic, which have thrived over the past few years by selling cutting-edge technology.

Perplexity this week previewed a new system for its computing product built around GLM 5.2, an open model from China’s Z.ai. The system is designed to allow a cheaper model to handle more of the job, calling out a more powerful model only when needed.

This approach reflects a broader shift in the market. Open-heavy models that can be downloaded, set up and run by the companies themselves are becoming more capable. They are also cheaper to run than premium proprietary models from the largest AI labs.

Peter Fenton, Benchmark’s general partner, said the change could be dramatic.

“An opposing view that perhaps reaches consensus is our belief that more than 90 percent of tokens created will exit gap-weighted models within the next 18 to 24 months, possibly even by the end of the year,” Fenton told CNBC.

Tokens are units of data processed and produced by artificial intelligence models.

“I think the inference margins that the leading model companies are producing are going to come under pressure when you have good enough models from explicit weights that you can run them without the markup that they provide,” Fenton said.

The switch to open models isn’t just about saving money, Fenton said. In some cases, smaller models tuned for a specific task can be faster and perform better than larger general-purpose models.

That’s one of the reasons why Benchmark invested in Ollama, a company that makes it easy for developers and organizations to download, run and manage open models.

“What matters is where the model comes from, where it was created and trained,” said Ollama CEO Jeff Morgan. “But what’s more important for these businesses we’re talking to is where and how they operate.”

Ollama has been adopted by more than 85% of the Fortune 500, including companies in regulated industries such as aviation, insurance and healthcare, Morgan said. Many companies start with smaller models that work closely with their own data, then expand to larger open models as they become more comfortable, he said.

The rise of open models also creates a strategic challenge for the U.S. Many of the most competitive open-heavy models come from Chinese labs, including Z.ai and Z.ai. DeepSeek. This has made open source AI a business issue, a policy issue, and a national competitive issue.